# TruePoint Claim Grounding

Pay-per-request claim verification for AI agents.

Base URL: https://claim-grounding.prospectintel.workers.dev
Protocol: x402
Network: Base mainnet (eip155:8453)
Asset: USDC
Pay to: 0x485FfC037f02dDDc154A454a9cA89C4B1b29e25b
Facilitator: Coinbase CDP
Version: 1.5.0
Sources: Wikipedia, Wikidata, Google News, Europe PMC (PubMed), FRED and SEC EDGAR. verify_claim uses Wikipedia; check_answer uses Wikipedia and Wikidata, plus Google News titles for recent or current claims; deep_verify uses all six.

Naming:
- POST /v1/deep_verify = Deep verify (not the paid research SKU)
- POST /v1/deep_research = Deep Research (confirmed spend only)
- budget.allow_deep does not authorize spend

Endpoints:
- GET /  service info
- GET /.well-known/agent-card.json
- GET /.well-known/x402
- GET /openapi.json
- GET /accuracy  published test results and limits
- POST /mcp  public remote MCP server (Streamable HTTP)
- POST /v1/quick_check      $0.15  fast supported/contradicted/unsure verdict on one claim, 1-2 sources
- POST /v1/extract_claims   $0.05  split any AI answer into checkable claims; up to 3 get a ready-to-send next_step (quick_check $0.15 first; deep_verify $0.75 / verify_claim $1.00)
- POST /v1/verify_claim     $1.00
- POST /v1/check_answer     $1.00
- POST /v1/deep_verify      $0.75
- POST /v1/deep_research/lite $5.00   Lite, requires confirmed: true
- POST /v1/deep_research    $8.00   Core, requires confirmed: true
- POST /v1/deep_research/pro  $15.00  Pro, requires confirmed: true
- GET /v1/jobs/:id  poll, no extra charge

How to use:
1. POST JSON to a paid endpoint.
2. If you receive HTTP 402, pay the x402 requirement and retry.
3. verdict is one of supported, contradicted, outdated, mixed, insufficient_evidence (legacy_verdict keeps the pre-1.2.0 value). On anything but supported, next_actions offers Deep Research. Call that route only with confirmed: true.
4. Bodies: extract_claims {"text"}, verify_claim {"claim":{"text"}}, check_answer {"answer"}, deep_verify {"claim":{"text"}}, deep_research {"confirmed":true,"answer"}.
5. Extract claims (1.4.7): each claim has claim_type and check_worthiness; hedges stripped; pronouns resolved; compounds split.
6. Check answer (1.4.7): Wikipedia plus Wikidata structured facts for dates, numbers, quantities, ordinals and records, and Google News titles only for recent or current claims (news never contradicts); returns summary (overall_verdict, counts, problem_sentences) and per-claim suggested_correction (quoted from cited sources only). Abstained when evidence is weak.
7. Your own sources (1.4.6): verify_claim, check_answer and deep_verify accept "sources": up to 10 items, each {"url":"https://..."} (public https page; fetched with 3.5 s timeout and 512 KB cap) or {"text":"...","title":"...","url":"..."} (up to 20,000 chars). Cited with source_type "caller"; reported in sources_used. "sources_only": true uses only your sources. Non-https, private/internal/metadata addresses, IP-literal hosts and custom ports are rejected with 400 (not charged). Results that use your sources are never cached.
6. MCP: public remote server (Streamable HTTP) at https://claim-grounding.prospectintel.workers.dev/mcp. tools/call returns x402 payment instructions; the paid HTTP route runs the work.
7. Accuracy results and what the tools do not do: https://claim-grounding.prospectintel.workers.dev/accuracy